Closed urialon closed 1 year ago
Adding self-consistency (majority@k).
On GSM, this improves PAL from 71.4 to 80.4 (majority@40). This makes PAL 2 points better than Minerva 540B when both models use majority@40.
LGTM!
Adding self-consistency (majority@k).
On GSM, this improves PAL from 71.4 to 80.4 (majority@40). This makes PAL 2 points better than Minerva 540B when both models use majority@40.